The browser extension is the attacker, the agent is the weapon
A browser extension on your personal laptop can get your work laptop to run code. That is the claim Gal Weizman of Forever Security put out on November 5 under the name ThisBash. The chain uses Claude Desktop's Dispatch feature, which exists to let a logged-in claude.ai web session send tasks to Claude Desktop on any machine linked to the same account. Weizman's research shows a browser extension with access to claude.ai can walk through that same door.
This is the second Forever Security disclosure in six weeks against the AI coding agent's own control surfaces. The first was EMBRACE on October 29 — the embedded browser pane inside five tools. ThisBash is the inter-device coordination channel. Two different surfaces, same shape.
The six steps
The Register writeup walks through the chain. A malicious extension that can run script on claude.ai does this:
- Registers a MAIN-world content script. MAIN-world scripts share the page's JavaScript context and session cookies — the extension is now indistinguishable from the logged-in user's own web tab.
- Opens Dispatch in a tiny, unfocused popup. The feature is loaded and ready, invisibly.
- Queries
GET /v1/sessions?tags=cowork-dispatch-localto retrieve the ID of the Claude Desktop agent session running on whatever machine is currently connected to this account. - Force-prompts the agent through the composer. The extension types instructions in the Dispatch composer. If the first phrasing does not succeed, it rephrases.
- Auto-clicks "Allow." The Claude Desktop agent waits for human approval before running sensitive actions. The extension clicks it for them, every second, until the agent moves forward.
- Repeats until code execution. Weizman's proof of concept had the agent run
npm install -gof a malicious package. The package'spreinstallhook opened Calculator.
Weizman told The Register the same underlying weakness exists in Claude Code Remote Control and Claude Cowork sessions.
The cross-device piece
Dispatch is explicitly cross-device. The attack inherits that design: a malicious extension on a personal browser session can reach a Claude Desktop instance on a company laptop as long as both are signed into the same Claude account. A user who installs a weather extension at home is not thinking about corporate access controls. The agent linked to their account is.
What was disclosed when
The Register and Dataconomy both report that Weizman reported the chain to Anthropic in May 2026. Anthropic validated it as real and closed the report as a duplicate of an issue it was already tracking. No CVE was issued, no bounty paid, and the feature was not patched. Anthropic had told users it planned to retire Dispatch at the end of October 2026; per the coverage, as of October 30 the feature remained available in limited beta on some Pro, Max, and Ultra plans. Weizman's public comment to The Register: "They still decided not to do anything about it."
Why this is the story
Prompt injection with no prompt channel to filter. Every step up to the final install is the extension driving the user's own authenticated session. The force-prompt is not a jailbreak buried in a document; the attacker is typing into the composer as the user. Any defense on the prompt channel is bypassed because the attacker holds the composer.
Human-in-the-loop defeat at the auto-click layer. The "Allow" dialog exists because agent tool calls are the thing the user should approve. If an attacker that already controls the page can also click on it, the dialog is UI, not enforcement.
A sequence, not a single step. MAIN-world injection, a documented endpoint call, typing into a composer, UI automation, an npm install. Each one is legitimate in isolation; the chain is the signal.
Two recent papers describe the same shape elsewhere: ETUDE at the tool-update layer, AuthorityFlow at the tool-argument layer. The November 4 survey Rethinking Access Control for AI Agents states the principle under all three: context is not authority. A browser-extension-controlled claude.ai session is context. It should not be authority to approve commands on a linked desktop.
Where Defend sits
This is the sequence shape Defend is built to watch. The individual primitives — a page opens Dispatch, a session endpoint is queried, the composer receives input, an "Allow" dialog is clicked, npm installs a package, a preinstall script executes — are all legitimate in isolation. The attack is their ordering and tempo. The force-prompt, the auto-approve, the install, the preinstall hook — that chain has no benign ordering, and watching the sequence is what catches it.
Vendor patches matter and we hope Anthropic ships one. The broader point — the same one EMBRACE made two weeks ago — is that the agent will keep acquiring new trust surfaces. The CVE catalog is not going to keep up.
What to do now
- Claude Desktop with Dispatch enabled: treat Dispatch as a privilege, not a convenience. If you do not need the feature, turn it off. If you do, do not stay signed into claude.ai in the same browser profile where you install extensions.
- If you use the same Claude account on both a personal and a corporate device: the per-account scoping is the attack surface. Separate accounts remove the cross-device primitive entirely.
- Browser extensions: audit what is loaded in the profile signed into claude.ai. A single compromised or acquired extension in that profile is the ThisBash initial access.
- Approval dialogs: any automation that can click them is an authentication-bypass primitive. Treat them as enforcement only when the click is a human click.
The agent is the weapon. The browser extension is the attacker. The control channel between them is the surface.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.